Amazon ECR containerd resolver

The Amazon ECR containerd resolver is an implementation of a containerd Resolver, Fetcher, and Pusher that can pull images from Amazon ECR and push images to Amazon ECR using the Amazon ECR API instead of the Docker Registry API.

Note: This repository is a proof-of-concept and is not recommended for production use.

Usage

Pull images

resolver, _ := ecr.NewResolver()
img, err := client.Pull(
    namespaces.NamespaceFromEnv(context.TODO()),
    "ecr.aws/arn:aws:ecr:us-west-2:123456789012:repository/myrepository:mytag",
    containerd.WithResolver(resolver),
    containerd.WithPullUnpack,
    containerd.WithSchema1Conversion)

Push images

ctx := namespaces.NamespaceFromEnv(context.TODO())

img, _ := client.ImageService().Get(
	ctx,
	"docker.io/library/busybox:latest")
resolver, _ := ecr.NewResolver()
err = client.Push(
	ctx,
	"ecr.aws/arn:aws:ecr:us-west-2:123456789012:repository/myrepository:mytag",
	img.Target,
	containerd.WithResolver(resolver))

Two small example programs are provided in the example directory demonstrating how to use the resolver with containerd.

ref

containerd specifies images with a ref. refs are different from Docker image names, as refs intend to encode an identifier, but not a retrieval mechanism. refs start with a DNS-style namespace that can be used to select separate Resolvers to use.

The canonical ref format used by the amazon-ecr-containerd-resolver is ecr.aws/ followed by the ARN of the repository and a label and/or a digest.

Parallel downloads

This resolver supports request parallelization for individual layers. This takes advantage of HTTP range requests to download different parts of the same file in parallel. This is an approach to achieving higher throughput when downloading from Amazon S3, which provides the raw blob storage for layers in Amazon ECR.

Request parallelization is not enabled by default, and the default Go HTTP client is used instead. To enable request parallelization, you can use the WithLayerDownloadParallelism resolver option to set the amount of parallelization per layer.

When enabled, the layer will be divided into equal-sized chunks (except for the last chunk) and downloaded with the set amount of parallelism. The chunks range in size from 1 MiB to 20 MiB; anything smaller than 1 MiB will not be parallelized and anything larger than 20 MiB * parallelism will use a larger number of chunks (though only with the specified amount of parallelism).

Initial testing suggests that a parallelism setting of 4 results in 3x faster layer downloads, but increases the amount of memory consumption between 15-20x. Further testing is still needed.

This support is backed by the htcat library.

Building

The Amazon ECR containerd resolver manages its dependencies with Go modules and requires Go 1.17 or greater. If you have Go 1.17 or greater installed, you can build the example programs with make.

License

The Amazon ECR containerd resolver is licensed under the Apache 2.0 License.